koharu: a local-first ML manga translator with automated inpainting and professional PSD export

What it solves

Koharu is a local-first manga translation tool that automates the tedious process of translating manga pages. It replaces the manual workflow of detecting text, removing original lettering, and typesetting translated text back into the page.

How it works

Koharu uses a staged pipeline of machine learning models to process pages:

1. Detection: Uses YOLO and other vision models to identify text regions and speech bubbles.
2. OCR: Recognizes the source text using multilingual OCR models.
3. Inpainting: Removes the original text from the image using models like FLUX.2 or LaMa to create a clean background.
4. Translation: Translates the text using either local LLMs (via llama.cpp) or remote APIs (OpenAI, Claude, Gemini, DeepSeek).
5. Rendering: A dedicated text renderer handles complex manga lettering, including vertical CJK layout and right-to-left scripts.

Who it’s for

It is designed for manga translators and scanlators who want a high-performance, privacy-focused desktop application that can run most of its pipeline locally on their own hardware.

Highlights

• Local-First Privacy: Runs vision models and LLMs locally to keep data secure.
• GPU Acceleration: Supports CUDA, Metal, Vulkan, and experimental ZLUDA for high-speed processing.
• Professional Export: Exports to layered PSD files, allowing for manual refinement of translated text layers in Photoshop.
• Agent Integration: Includes a built-in MCP server and headless mode for automation via local agents.
• Advanced Typesetting: Specialized rendering for vertical CJK and RTL scripts with Google Fonts support.